add IBM f128 type by folkertdev · Pull Request #162156 · rust-lang/rust

folkertdev · 2026-09-01T20:56:34Z

tracking issue: #161787

aka __ibm128, ppc_fp128, etc. Making this change without having settled on a final name is a bit awkward but hopefully we can get some rounds of review in anyway.

folkertdev · 2026-09-21T22:33:03Z

+    #[inline]
+    fn eq(&self, other: &ppcf128) -> bool {
+        self.to_components() == other.to_components()
+    }


Equality and comparison is just defined on the components in the obvious way

https://github.com/gcc-mirror/gcc/blob/c4a80f2d336b8fb78cc97f9107981119b1eff30b/libgcc/config/rs6000/ibm-ldouble.c#L280-L287

It's unclear to me why these are libcalls at all (you're not saving much code really).

View changes since the review

Neither Clang nor GCC seem to actually call these libcalls; I think the advantage of using the LLVM intrinsic is that it would enable further LLVM optimisations, but that definitely isn't needed for this initial PR.

folkertdev · 2026-09-21T22:34:06Z

+#[unstable(feature = "powerpc_ppcf128", issue = "161787")]
+#[allow(non_camel_case_types)]
+#[doc(cfg(any(target_arch = "powerpc", target_arch = "powerpc64")))]
+pub struct ppcf128(u128);


we may need/want to make this a #[align(16)] struct ppcf128([u8; 16]) instead to work around LLVM bitcast issues between i128 and ppcf128. We'll see.

View changes since the review

Wouldn't it make the most sense for it to be #[repr(C, align(16))] struct ppcf128(f64, f64)?

Right, or [f64; 2]. repr(C) is not quite right though (neither is repr(Rust) I guess, but the array solves the field ordering problem)

I opted for [f64; 2] because that is what Simd does too so it's less likely to cause issues.

Sadly that runs into some asserts (when debug_assertions are on) on Scalars truly being scalar (and not a wrapper around some other BackendRepr). So, I've reverted to u128 for now. I don't think LLVM can currently see that in the syntax we're using u128, and we're careful to transmute to [f64; 2], never u128, so we're probably fine.

In any case, it's not really something I want to block the initial PR on.

folkertdev · 2026-09-21T22:36:00Z

r? tgross35

rustbot · 2026-09-21T22:36:12Z

Some changes occurred in compiler/rustc_attr_ir

cc @jdonszelmann, @JonathanBrouwer

stdarch is developed in its own repository. If possible, consider making this change to rust-lang/stdarch instead.

cc @Amanieu, @sayantn

This PR changes rustc_public

cc @oli-obk, @celinval, @ouz-a, @makai410

rustc_codegen_gcc is developed in its own repository. If possible, consider making this change to rust-lang/rustc_codegen_gcc instead.

cc @antoyo, @GuillaumeGomez

rustc_codegen_cranelift is developed in its own repository. If possible, consider making this change to rust-lang/rustc_codegen_cranelift instead.

cc @bjorn3

beetrees · 2026-09-22T00:45:37Z

+        } else if large.is_infinite() && is_elfv2 {
+            small == 0.0


32-bit PowerPC also requires small == 0.0 per it's ABI spec. While the ELFv1 ABI spec doesn't explicitly mention that small == 0.0, if small != 0.0 then expressions like x == ppcf128::INFINITY won't work (as compilers just compare both parts without ignoring the second if the first is infinity), so it seems likely it is de-facto required even though not explicitly stated. Therefore I think it's safest to always require that small == 0.0 for now and check with LLVM and GCC what their expectations are before stabilisation.

View changes since the review

beetrees · 2026-09-22T00:47:00Z

+        } else if large.is_infinite() && is_elfv2 {
+            small == 0.0
+        } else {
+            large.abs() > small.abs() && large + small == large


Suggested change

large.abs() > small.abs() && large + small == large

large + small == large

large + small == large guarantees that large.abs() > small.abs().

View changes since the review

tgross35 · 2026-09-22T11:35:29Z

+
+    /// Negative infinity (−∞).
+    #[unstable(feature = "powerpc_ppcf128", issue = "161787")]
+    pub const NEG_INFINITY: Self = Self::from_components(f64::NEG_INFINITY, 0.0);


Could library support other than the type itself be split to a separate PR? There's enough to review on the compiler side and up to the codegen tests doesn't need any of this.

View changes since the review

Sure, I've split it out.

tgross35 · 2026-09-22T11:36:08Z

Should probably have an asm test to go with it

View changes since the review

Sure, I've added one.

tgross35 · 2026-09-22T11:41:15Z

        /// it can be safely ignored by always picking i8.
        hint_vector_elem: Primitive,
    },
 }


This feels like an unusual addition since the other variants in this enum are scalar registers. Can we treat it more like a scalarpair or a struct { f64, f64 } for Rust's ABI purposes?

@RalfJung may have ideas

View changes since the review

I don't think that works, if you look at where this is needed (specifically in homogenous_aggregate).

Note that f16b actually runs into the same issue, I've just not been able to get it to miscompile but in aggregates bfloat turns into half, e.g. https://godbolt.org/z/Mxr7axj7z.

The entire callconv system is terrible and needs to be redesigned from scratch. We're just making the inevitable rewrite harder by continuing to delay it. Now that we have people funded to work on rustc I hope we can get those funds to be used for less fun work such as a rewrite. :)

(Actually to me that sounds like very fun work. If I had the time I'd do it. But it's unlikely that I will have time time any time soon.)

In the mean time we can continue to add more hacks to the existing hacks. 🤷

RegKind at the moment is more "primitive type passed to LLVM", so adding PpcF128 here isn't making the problem any worse (e.g. a RegKind::Integer can get split over multiple actual integer registers).

That seems worth putting in a doc comment on this type.

(To be slightly more precise, RegKind is the "kind" of LLVM type, which is combined with a Size to make Reg, which represents the "primitive type passed to LLVM".)

I've added some brief comments (trying to use "backend" instead of "llvm").

I think LLVM would be more honest here. ;) Other backends just somehow have to deal with this.

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch from d0bb095 to 8b5b4be Compare September 2, 2026 13:03

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch from 8b5b4be to 669f1f0 Compare September 2, 2026 13:20

rustbot added the T-clippy Relevant to the Clippy team. label Sep 2, 2026

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch from 669f1f0 to 802c214 Compare September 2, 2026 20:16

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch from 802c214 to 39c2bba Compare September 3, 2026 17:33

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch 2 times, most recently from 2ef8b4a to fbe3899 Compare September 21, 2026 17:11

powerpc is_homogeneous_aggregate cleanup

961481d

folkertdev force-pushed the ppcf128 branch from fbe3899 to 66d3dfb Compare September 21, 2026 18:29

folkertdev added the F-c_longdouble `#![feature(c_longdouble)]` label Sep 21, 2026

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch from 66d3dfb to 5f7f25a Compare September 21, 2026 18:46

This comment has been minimized.

Sign in to view

folkertdev force-pushed the ppcf128 branch 2 times, most recently from d0e13f1 to 9251025 Compare September 21, 2026 22:01

folkertdev commented Sep 21, 2026

View reviewed changes

rustbot assigned tgross35 Sep 21, 2026

folkertdev marked this pull request as ready for review September 21, 2026 22:36

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Sep 21, 2026

beetrees reviewed Sep 22, 2026

View reviewed changes

Comment thread compiler/rustc_abi/src/callconv/reg.rs Outdated

Comment thread compiler/rustc_codegen_llvm/src/llvm/ffi.rs Outdated

folkertdev force-pushed the ppcf128 branch from 9251025 to c6f8c21 Compare September 22, 2026 00:23

beetrees reviewed Sep 22, 2026

View reviewed changes

tgross35 reviewed Sep 22, 2026

View reviewed changes

add Float::PpcF128

40633d3

folkertdev force-pushed the ppcf128 branch from c6f8c21 to abc4208 Compare September 22, 2026 13:47

This comment has been minimized.

Sign in to view

folkertdev added 5 commits September 22, 2026 17:16

add core::arch::powerpc::ppcf128

2bd8d88

add RegKind::PpcF128

9cd3fe3

add ppcf128 abi test

55b1d75

add ppcf128 assembly test

c5e25a1

add doc comments on Reg and RegKind

5d87632

folkertdev force-pushed the ppcf128 branch from abc4208 to 5d87632 Compare September 22, 2026 15:31

	large.abs() > small.abs() && large + small == large
	large + small == large

Uh oh!

Conversation

folkertdev commented Sep 1, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

folkertdev Sep 21, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

folkertdev Sep 21, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

folkertdev commented Sep 21, 2026

Uh oh!

rustbot commented Sep 21, 2026

Uh oh!

Uh oh!

Uh oh!

beetrees Sep 22, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

beetrees Sep 22, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tgross35 Sep 22, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tgross35 Sep 22, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tgross35 Sep 22, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

RalfJung Sep 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

beetrees Sep 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

folkertdev commented Sep 1, 2026 •

edited by rustbot

Loading

folkertdev Sep 21, 2026 •

edited by rustbot

Loading

folkertdev Sep 21, 2026 •

edited by rustbot

Loading

beetrees Sep 22, 2026 •

edited by rustbot

Loading

beetrees Sep 22, 2026 •

edited by rustbot

Loading

tgross35 Sep 22, 2026 •

edited by rustbot

Loading

tgross35 Sep 22, 2026 •

edited by rustbot

Loading

tgross35 Sep 22, 2026 •

edited by rustbot

Loading

RalfJung Sep 22, 2026 •

edited

Loading

beetrees Sep 22, 2026 •

edited

Loading